Research map / Architectures and efficiency / RNNs and normalization
Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling
Qiao Xiao, Boqian Wu, Patrik Okanovic, Tomasz Sternal, Maurice van Keulen, Elena Mocanu, Mykola Pechenizkiy, Decebal Constantin Mocanu, Torsten Hoefler
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add optimizer-state warm-up to Adam-based DST (ablation) | SET + Adam · C4 LLaMA-350M ablation (d=0.1) | perplexity: 28.98 -> 27.67 | Improved |
| add density-aware LR scaling to Adam-based DST (ablation) | SET + Adam · C4 LLaMA-350M ablation (d=0.1) | perplexity: 28.98 -> 28.20 | Improved |
| add LR scaling to warm-up (full SMET) (ablation) | SMET · C4 LLaMA-350M ablation (d=0.1) | perplexity: 27.67 -> 25.85 | Improved |
| add warm-up to LR scaling (full SMET) (ablation) | SMET · C4 LLaMA-350M ablation (d=0.5) | perplexity: 24.79 -> 21.61 | Improved |
| use warm-up length W=10 (ablation) | SMET · C4 LLaMA-350M density 0.2 | perplexity | Improved |
| use topology update interval of 100 steps instead of 50 or 300 (ablation) | SMET · C4 LLaMA-350M | perplexity | Improved |
| replace SET with SMET (random regrowth) | SET · C4 sparse pre-training (d=0.1) | perplexity: 28.96 -> 25.85 | Improved |
| replace RigL with SMET_grad (gradient regrowth) | RigL · C4 sparse pre-training (d=0.1) | perplexity: 29.90 -> 26.50 | Improved |
| dynamic sparse training with RigL versus static sparse | Static Sparse · C4 sparse pre-training (d=0.5) | perplexity: 22.90 -> 25.75 | Hurt |
| SMET versus static sparse | Static Sparse · C4 sparse pre-training (d=0.5) | perplexity: 22.90 -> 21.61 | Improved |
| SMET versus GaLore at 1B | GaLore · C4 LLaMA-1B | perplexity: 15.64 -> 15.61 | No clear effect |
| SMET versus GaLore at 350M | GaLore · C4 LLaMA-350M | perplexity: 18.95 -> 18.90 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ReLoRA — ReLoRA: High-Rank Training Through Low-Rank Updates (ICLR 2024) · better on 4 of 4 shared results
- GaLore — GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.91)
- Navigating Extremes: Dynamic Sparsity in Large Output Spaces (NeurIPS 2024 · similarity 0.91)
- Advancing Dynamic Sparse Training by Exploring Optimization Opportunities (ICML 2024 · similarity 0.89)
- Mini-batch Coresets for Memory-efficient Language Model Training on Data Mixtures (ICLR 2025 · similarity 0.89)
- R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference (ICLR 2025 · similarity 0.88)
- Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint? (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
